Weights animation: sliding-window style in a fixed space

Switch back from animate='spin' to the classic animate=True sliding
window, per review. The critical property -- the space inside the cube
is fixed for the whole animation, independent of which window is
visible -- is guaranteed by the pipeline (helpers.scale normalizes once
from the full stacked dataset and the window updater never touches axis
limits) and verified programmatically: axis limits are identical across
frames while the visible fragment's extent slides along the loop.

(The earlier claim that window mode showed a static tangle was a frame-
extraction bug in the verification harness -- PIL's ImageSequence
yields one re-seeked image object, so materializing it with list()
produced N copies of the final frame. Measured correctly, the window
render rotates and the comet travels: mean inter-second frame motion
8.3, 0 clipped frames.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: jeremymanning

docs/images/v2.0-animations/weights_hyperaligned.gif

@@ -150,7 +150,8 @@ Other round-4.5 fixes: repeated-hyperalign scale collapse (procrustes optimal sc
 
 ## Round 4.5 SHIPPED (2317cd5), then weights loop-shape correction (2026-07-02 late)
 - WEIGHTS FINAL recipe (scripts/generate_weights_trajectory.py): smooth(gaussian var=300, scipy gaussian_filter1d sigma=sqrt(300)) -> align('SRM', n_iter=20) -> smooth -> reduce={'model':'UMAP','params':{'n_neighbors':36,'min_dist':0.1,'random_state':42}} -> **animate='spin'** (NOT window), duration=30, frame_rate=30, rotations=1, linewidth=3, zoom=2, size=[8,6].
-  - CORRECTION vs 2317cd5: had shipped n_neighbors=150; Jeremy: that gives "a straight (slightly curved) line, with no dramatic bends" — download.png has a LOOP. Swept nn={15,30,36,50,80,150} x min_dist x seeds vs download.png. nn=15 hairball; nn=150 over-globalizes -> flattens loop; **nn=36 md=0.1 = tight bundle WITH the loop** (best qualitative match). Also switched window animation (animate=True + tail_duration=4) -> animate='spin': window only ever showed a rolling 4s fragment (tangle), never the full loop; spin draws the whole bundle and orbits so the loop is always visible.
+  - CORRECTION vs 2317cd5: had shipped n_neighbors=150; Jeremy: that gives "a straight (slightly curved) line, with no dramatic bends" — download.png has a LOOP. Swept nn={15,30,36,50,80,150} x min_dist x seeds vs download.png. nn=15 hairball; nn=150 over-globalizes -> flattens loop; **nn=36 md=0.1 = tight bundle WITH the loop** (best qualitative match).
+  - ANIMATION STYLE (Jeremy, round 4.6): sliding window (animate=True + tail_duration=4), NOT spin (briefly shipped spin in b362d67; reverted). CRITICAL REQUIREMENT: the view/space inside the cube must be FIXED across the full animation — must not depend on which window is visible. Already guaranteed by implementation: helpers.scale() normalizes once from the FULL stacked data (plot.py:614) and update_lines_parallel never touches limits; VERIFIED programmatically (limits identical across frames: +/-1.15/+/-1.04, while fragment x-extent slides -0.99..0.04 -> 0.36..1.00 -> -0.80..-0.11).
   - PROVED it's a pure UMAP-n_neighbors effect, not a version/alignment artifact: (a) modern SRM branch align.py:137 SRM(features=min shape[0]) re-fit per pass is byte-identical to era 0.6.2 align.py SRM branch; (b) built era venv (uv, py3.9, numba 0.55, umap-learn 0.4.6, patched layouts.py int32->intp) and fit 0.4.6 on the SAME modern-aligned data -> ALSO a hairball at its default nn=15. So old deps don't recover the look; neighborhood tuning does. Byte-exact era repro infeasible on arm64 (ht 0.6.2 pins sklearn 0.21.3, won't build). download.png = DESIRED look, acceptance qualitative (bundle + loop, not a straight line). Cached: /tmp/aligned_weights.npz (modern SRM x20 aligned), /tmp/raw_weights.npz.
   - SRM corr plateaus 0.87 (shared-signal ceiling; SRM unchanged since v0.6.2).
 - align('SRM') n_iter support; hyperalign per-pass rescale (collapse fix).

notes/session_2026-07-01_dev-2.0_kickoff.md

@@ -52,10 +52,12 @@ def main():
              reduce={'model': 'UMAP',
                      'params': {'n_neighbors': 36, 'min_dist': 0.1,
                                 'random_state': 42}},
-             # spin (not window) so the complete looping bundle stays on
-             # screen and orbits in 3D -- the classic readthedocs gif style;
-             # a rolling window would only ever show a tangled fragment
-             animate='spin', duration=30, frame_rate=30,
+             # sliding-window animation in a FIXED space: the data are
+             # scaled once from the full dataset (helpers.scale), so the
+             # axis limits never depend on which window is visible -- the
+             # comet travels along the loop while the cube's contents stay
+             # put (verified: limits identical across frames)
+             animate=True, duration=30, frame_rate=30, tail_duration=4,
              rotations=1, linewidth=3, zoom=2, size=[8, 6],
              save_path=mp4, show=False)
     # the dense spinning bundle (36 fat trajectories every frame) is large;